ACTA UNIVERSITATIS LODZIENSIS FOLIA OECONOMICA 286, 2013 Katarzyna Dębkowska *, Marta Jarocka ** THE IMPACT OF THE METHODS OF THE DATA NORMALIZATION ON THE RESULT OF LINEAR ORDERING Abstract. In taxonomy, various methods of the data normalization are used. They can significantly affect the result of the classification. This article presents the analysis of the impact of the methods of the data normalization on the result of linear ordering. The following methods and tools are used in this research: standardization, Weber s standardization, unitization, zero unitization with zero minimum and selected quotient transformation. A statement comparing the methods used in the study of classification is formulated. In the research data concerning innovation of European Union regions from Eurostat are used. Key words: normalization, linear ordering. I. INTRODUCTION Normalization of the variables is a problem which is often considered in the literature in the field of multidimensional statistics. Most information about the normalization of the variables can be found in scientific papers containing the results of empirical comparative studies of complex economic phenomena that are preceded by a theoretical part describing the methodology used for supplying data for comparability. Detailed and in-depth discussion of normalization can be found in few works. The studies which are noteworthy in this field are K. Kukuła (2000) and B. Pawelek (2008). This article presents the results of a research experiment, which aims to analyze the impact of various normalization procedures on the result of linear ordering of obects. The paper is based on normalization formulas including methods such as standardization, Weber s standardization, unitization, zero unitization and selected transformation quotients. Linear ordering was performed by means of model as well as non model methods. Results are presented on the empirical example considering linear ordering on EU regions because of their innovative features. * Ph.D., Chair of Business Informatics and Logistics, Bialystok University of Technology. ** Master, Chair of Business Informatics and Logistics, Bialystok University of Technology. [181]
182 II. NORMALIZATION OF THE VARIABLES AS ONE OF THE STAGES OF LINEAR ORDERING Normalization of the variables is one of the stages of linear ordering, and its task is to deprive titers features of their value and to unify the range of magnitude in order to bring them to comparability. The research experiment sought the answers to such questions as: how does changing normalization transformation formula affect the outcome of linear ordering, as well as which of the normalization methods give similar results of linear ordering. The most commonly used normalization formulas were presented in Table 1. It also contains information on measurement scales of the input and output variables. The type of the scale of the input variables determines the set of acceptable normalization transformations (Jauga, Walesiak, 2000), (Pawelek, 2008), (Walesiak, 2011). Table 1. Types of variable normalization formulas The scale of measurement Sign. Name Formula variables before after normalization normalization S standardization z ( x x ) / s ratio scale and (or) interval interval SW Weber z ( x Me ) /1. 4826MAD ratio scale and standardization (or) interval interval U unitization z ( x x )/ r ratio scale and (or) interval interval UZ unitization with z [ x min{ x}) / r ratio scale and zero minimum (or) interval interval max z x / max{ x } ratio scale ratio scale s r x sum psk quotient transformation i z x / s ratio scale ratio scale z x / r ratio scale ratio scale z x / x ratio scale ratio scale z z n x i 1 n 2 x i 1 x / ratio scale ratio scale x / ratio scale ratio scale Source: M. Walesiak, Uogólniona miara odległości GDM w statystyczne analizie wielowymiarowe z wykorzystaniem programu R, Wyd. UE we Wrocławiu, 2011, p. 19.
The Impact of the Methods of the Data Normalization 183 When deciding on normalization methods, besides the variables measurement scales, we should take into account the descriptive characteristics of the input diagnostic variables (Pawelek, 2008) and characteristics of the distribution of the variables, designated for normalized variable values (Walesiak 2011). Normalization formulas such as unitization, zero unitization, ratio scale transformation with normalization basis which is equal to range are valuable because they provide normalized values of the variables varied variability (measured by standard deviation) and at the same time constant range for all variables. Classic standardization, Weber s standardization and quotient transformation with the basis of normalization equal to standard deviation, cause unification of the variables in terms of variability which is measured by means of standard deviation (absolute median deviation) this means the elimination of variation as a basis for differentiating obects. Quotient transformations with the normalization basis equal to maximum and square root of the sum of squared observations provide a diverse: variability, arithmetic mean and range for normalized variable values. Quotient transformations with the normalization basis equal to the sum and the arithmetic mean provide a diverse variability, range and constant arithmetic mean for normalized variables value. All normalization formulas, being the linear observation transformations of each variable, keep the skewness and kurtosis of the variables distribution and for each pair of variables all normalization formulas do not change the value of the Pearson s correlation coefficient. III. EXPERIMENTAL RESEARCH RESULTS The study involved 217 EU regions due to their innovative features. Preceded by a statistical analysis, the selection of diagnostic features was as follows: Personnel in the research and development sector as a percentage of total employment; EPO (European Patent Office) patent applications per million inhabitants; Expenditure on research and development per 1 inhabitant in the business sector (EURO per 1 inhabitant); Employment in science and technology as a percentage of the economically active population. Realizations of diagnostic features were normalized according to different normalization formulas and next was performed a linear ordering of regions using non model method by means of arithmetic mean. Table 2 shows the positions of the most innovative regions in the rankings obtained by using different normalization formulas.
184 Table 2. The positions of regions in the rankings (linear ordering: non model method by means of arithmetic mean) Signature of normalization formulas Region s symbol S SW U UZ s r x max sum psk DE11 1 1 1 1 1 1 1 1 1 1 DK01 2 3 2 2 2 2 2 2 2 2 DE21 3 2 3 3 3 3 3 3 3 3 SE11 4 4 4 4 4 4 4 4 4 4 DE14 5 5 5 5 5 5 5 5 5 11 DE71 6 8 8 8 8 6 8 10 10 12 DE91 7 9 6 6 6 7 6 6 6 6 DE12 8 7 7 7 7 8 7 7 7 16 FI18 9 13 9 9 12 9 9 14 14 18 FI1A 10 11 10 10 9 10 10 11 11 9 The data presented in Table 2 shows that the change in the normalization procedure affects the change of position in the analyzed region rankings. We can see them even in the top ten most innovative regions, for instance, regions marked by symbols DE71, DE12, FI18 changed their positions in the presented classifications even by a few places. In order to compare the rankings obtained in the research, Spearman correlation coefficients were calculated, whose values fluctuate in the range of 0.9624 1.0000 (Table 3). Table 3. Correlation matrix between result s rankings obtained by using different normalization formulas (linear ordering: non model method by means of arithmetic mean) Formula s sign. S SW U UZ max s r x sum psk S 1.000 0.9936 0.9997 0.9997 0.9986 1.0000 0.9997 0.9819 0.9819 0.9624 SW 1.0000 0.9946 0.9946 0.9961 0.9936 0.9946 0.9947 0.9947 0.9796 U 1.0000 1.0000 0.9994 0.9997 1.0000 0.9843 0.9843 0.9631 UZ 1.0000 0.9994 0.9997 1.0000 0.9843 0.9843 0.9631 max 1.0000 0.9986 0.9994 0.9886 0.9886 0.9669 s 1.0000 0.9997 0.9819 0.9819 0.9624 r 1.0000 0.9843 0.9843 0.9631 x 1.0000 1.0000 0.9821 sum 1.0000 0.9821
The Impact of the Methods of the Data Normalization 185 Despite a high degree of correlation of linear ordering results, by means of different normalization methods we can observe significant differences in the results of particular rankings. Table 4 presents the number of regions whose positions have changed in relation to their position in the ranking obtained by means of standardization. These numbers correspond to the number of "shifts" of regions positions in the analyzed rankings. Table 4. Change in the positions of regions in the rankings (linear ordering: non model method by means of arithmetic mean) Number of "shifts" position in the ranking of regions relative to ranking with standardization Number of regions whose position has changed relative to ranking using standardization formula SW U UZ s r x max sum psk 0 16 69 69 32 217 69 8 8 15 1 44 87 87 62 0 87 18 18 18 2 36 43 43 48 0 43 23 23 17 3 18 8 8 25 0 8 15 15 16 4 13 8 8 13 0 8 16 16 13 5 16 1 1 14 0 1 16 16 13 6 12 0 0 9 0 0 21 21 9 7 11 0 0 6 0 0 12 12 13 8 10 0 0 3 0 0 7 7 8 9 5 0 0 2 0 0 6 6 6 10 7 2 2 1 0 2 7 7 3 29 2 0 0 0 0 0 0 0 3 83 0 0 0 0 0 0 0 0 1 The uxtaposition presented in Table 4 shows that with the change in the normalization procedure, positions of evaluated regions have changed as well. The biggest changes occurred in the case of replacement of the classic standardization by Weber s standardization or by selected quotient transformations. For example, a change of data normalization method from the standardization to Weber s standardization caused the fact that the position of two regions have shifted in the ranking list by 29 places, and when using quotient transformation psk the position of one of the regions changed up to 83 places.
186 Identical calculations were performed in the next stage of the research experiment, replacing the non model linear ordering method by parametric Hellwig method (Hellwig, 1968). Tables 5, 6 and 7 present the results of the calculations which are equivalent to the results described in Tables 2, 3 and 4. Table 5. The positions of regions in the rankings (linear ordering: parametric Hellwig method) Region s symbol S SW U UZ Signature of normalization formulas s r x max sum DE11 1 1 1 1 1 1 1 1 1 1 DE21 2 2 2 2 2 2 2 2 2 5 DK01 3 4 3 3 3 3 3 3 3 2 SE11 4 3 4 4 4 4 4 4 4 3 DE71 5 6 5 5 6 5 5 7 7 11 SE22 6 8 8 8 8 6 8 6 6 8 DE14 7 5 6 6 5 7 6 5 5 14 DE12 8 7 7 7 7 8 7 8 8 21 FI18 9 12 9 9 10 9 9 15 15 16 DE91 10 11 10 10 9 10 10 9 9 6 psk Table 6. Correlation matrix between result s rankings obtained by using different normalization formulas (linear ordering: parametric Hellwig method) Formula s sign. S SW U UZ max s r x sum psk S 1.000 0.981 0.999 0.999 0.997 1.000 0.999 0.970 0.970 0.946 SW 1.000 0.983 0.983 0.987 0.981 0.983 0.996 0.996 0.968 U 1.000 1.000 0.999 0.999 1.000 0.973 0.973 0.947 UZ 1.000 0.999 0.999 1.000 0.973 0.973 0.947 max 1.000 0.997 0.999 0.981 0.981 0.955 s 1.000 0.999 0.970 0.970 0.946 r 1.000 0.973 0.973 0.947 x 1.000 1.000 0.979 sum 1.000 0.979
The Impact of the Methods of the Data Normalization 187 Table 7. Change in the positions of regions in the rankings (linear ordering: parametric Hellwig method) Number of regions whose position has changed relative to ranking using standardization formula Number of "shifts" position in the ranking of regions relative to ranking with standardization SW U UZ max s r x sum 0 13 53 53 27 217 53 11 11 10 1 16 61 61 31 0 61 15 15 10 2 22 37 37 35 0 37 14 14 14 3 21 36 36 27 0 36 16 16 9 4 10 15 15 25 0 15 11 11 17 5 10 8 8 17 0 8 11 11 8 6 16 4 4 22 0 4 7 7 7 7 9 2 2 7 0 2 8 8 6 8 12 0 0 9 0 0 7 7 8 9 10 2 2 6 0 2 7 7 7 10 5 0 0 2 0 0 13 13 8 35 2 0 0 0 0 0 1 1 2 50 2 0 0 0 0 0 1 1 0 60 0 0 0 0 0 0 1 1 1 psk Using the model linear ordering method leads to the same conclusions as in the case of application of non- model procedure. Changing the normalization method affects the positions of obects in the rankings, regardless of the classification procedure. IV. SUMMARY Summarizing the above considerations it must be explicitly stated that the choice of normalization formula influences the result of linear ordering, in both model and non model methods. By the realization of a research experiment it was noted that, despite high levels of numerical correlation coefficients calculated between places in the rankings determined by using different methods of normalization, we can observe changes in the positions in the rankings in case of particular regions. Therefore, it is recommended to use the same normalization procedure of diagnostic variables while creating periodical classifications used to observe obects in time.
188 ACKNOWLEDGMENT This scientific research was co-financed from the resources of the National Centre for Science nr N N111 530140. REFERENCES Hellwig Z. (1968), Zastosowanie metody taksonomiczne do typologicznego podziału kraów ze względu na poziom ich rozwou oraz zasoby i strukturę wykwalifikowanych kadr, Przegląd Statystyczny, nr 4, pp. 307-326. Kukuła K. (2000), Metoda unitaryzaci zerowe, Wydawnictwo Naukowe PWN, Warszawa. Pawełek B. (2008), Metody normalizaci zmiennych w badaniach porównawczych złożonych zawisk ekonomicznych, Wyd. Uniwersytetu Ekonomicznego w Krakowie. Walesiak M. (2011), Uogólniona miara odległości GDM w statystyczne analizie wielowymiarowe z wykorzystaniem programu R, Wyd. UE we Wrocławiu. Jauga K., Walesiak M. (2000), Standardization of Data Set under Different Measurement Scales [w:] Classification and Information Processing AT the Turn of the Millennium, red. R. Decker, W. Gaul, Springer Verlag, Berlin-Heidelberg. WPŁYW FORMUŁY NORMALIZACYJNEJ NA WYNIK PORZĄDKOWANIA LINIOWEGO W taksonomicznych badaniach empirycznych spotyka się różne formuły transformaci normalizacyne. Często poma się fakt, iż e wybór może istotnie wpływać na wynik klasyfikaci. W artykule zaprezentowano rezultaty eksperymentu badawczego, którego celem była analiza wpływu różnych procedur normalizacynych na wynik porządkowania liniowego obiektów. W pracy wykorzystano między innymi takie formuły normalizacyne ak: standaryzaca, standaryzaca Webera, unitaryzaca, unitaryzaca zerowa oraz wybrane przekształcenia ilorazowe. Następnie dokonano komparaci otrzymanych wyników. Procedurę badawczą oparto o dane zaczerpnięte z bazy Eurostat, które dotyczyły innowacyności regionów Unii Europeskie. Zawisko innowacyności regionów zostało scharakteryzowane poprzez szereg cech, których realizace były dostępne na poziomie badanych ednostek terytorialnych.